• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* origin: OpenBSD /usr/src/lib/libm/src/ld80/e_powl.c */
2 /*
3  * Copyright (c) 2008 Stephen L. Moshier <steve@moshier.net>
4  *
5  * Permission to use, copy, modify, and distribute this software for any
6  * purpose with or without fee is hereby granted, provided that the above
7  * copyright notice and this permission notice appear in all copies.
8  *
9  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
10  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
11  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
12  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
13  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
14  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
15  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
16  */
17 /*                                                      powl.c
18  *
19  *      Power function, long double precision
20  *
21  *
22  * SYNOPSIS:
23  *
24  * long double x, y, z, powl();
25  *
26  * z = powl( x, y );
27  *
28  *
29  * DESCRIPTION:
30  *
31  * Computes x raised to the yth power.  Analytically,
32  *
33  *      x**y  =  exp( y log(x) ).
34  *
35  * Following Cody and Waite, this program uses a lookup table
36  * of 2**-i/32 and pseudo extended precision arithmetic to
37  * obtain several extra bits of accuracy in both the logarithm
38  * and the exponential.
39  *
40  *
41  * ACCURACY:
42  *
43  * The relative error of pow(x,y) can be estimated
44  * by   y dl ln(2),   where dl is the absolute error of
45  * the internally computed base 2 logarithm.  At the ends
46  * of the approximation interval the logarithm equal 1/32
47  * and its relative error is about 1 lsb = 1.1e-19.  Hence
48  * the predicted relative error in the result is 2.3e-21 y .
49  *
50  *                      Relative error:
51  * arithmetic   domain     # trials      peak         rms
52  *
53  *    IEEE     +-1000       40000      2.8e-18      3.7e-19
54  * .001 < x < 1000, with log(x) uniformly distributed.
55  * -1000 < y < 1000, y uniformly distributed.
56  *
57  *    IEEE     0,8700       60000      6.5e-18      1.0e-18
58  * 0.99 < x < 1.01, 0 < y < 8700, uniformly distributed.
59  *
60  *
61  * ERROR MESSAGES:
62  *
63  *   message         condition      value returned
64  * pow overflow     x**y > MAXNUM      INFINITY
65  * pow underflow   x**y < 1/MAXNUM       0.0
66  * pow domain      x<0 and y noninteger  0.0
67  *
68  */
69 
70 #include "libm.h"
71 
72 #if LDBL_MANT_DIG == 53 && LDBL_MAX_EXP == 1024
powl(long double x,long double y)73 long double powl(long double x, long double y)
74 {
75 	return pow(x, y);
76 }
77 #elif LDBL_MANT_DIG == 64 && LDBL_MAX_EXP == 16384
78 
79 /* Table size */
80 #define NXT 32
81 
82 /* log(1+x) =  x - .5x^2 + x^3 *  P(z)/Q(z)
83  * on the domain  2^(-1/32) - 1  <=  x  <=  2^(1/32) - 1
84  */
85 static const long double P[] = {
86  8.3319510773868690346226E-4L,
87  4.9000050881978028599627E-1L,
88  1.7500123722550302671919E0L,
89  1.4000100839971580279335E0L,
90 };
91 static const long double Q[] = {
92 /* 1.0000000000000000000000E0L,*/
93  5.2500282295834889175431E0L,
94  8.4000598057587009834666E0L,
95  4.2000302519914740834728E0L,
96 };
97 /* A[i] = 2^(-i/32), rounded to IEEE long double precision.
98  * If i is even, A[i] + B[i/2] gives additional accuracy.
99  */
100 static const long double A[33] = {
101  1.0000000000000000000000E0L,
102  9.7857206208770013448287E-1L,
103  9.5760328069857364691013E-1L,
104  9.3708381705514995065011E-1L,
105  9.1700404320467123175367E-1L,
106  8.9735453750155359320742E-1L,
107  8.7812608018664974155474E-1L,
108  8.5930964906123895780165E-1L,
109  8.4089641525371454301892E-1L,
110  8.2287773907698242225554E-1L,
111  8.0524516597462715409607E-1L,
112  7.8799042255394324325455E-1L,
113  7.7110541270397041179298E-1L,
114  7.5458221379671136985669E-1L,
115  7.3841307296974965571198E-1L,
116  7.2259040348852331001267E-1L,
117  7.0710678118654752438189E-1L,
118  6.9195494098191597746178E-1L,
119  6.7712777346844636413344E-1L,
120  6.6261832157987064729696E-1L,
121  6.4841977732550483296079E-1L,
122  6.3452547859586661129850E-1L,
123  6.2092890603674202431705E-1L,
124  6.0762367999023443907803E-1L,
125  5.9460355750136053334378E-1L,
126  5.8186242938878875689693E-1L,
127  5.6939431737834582684856E-1L,
128  5.5719337129794626814472E-1L,
129  5.4525386633262882960438E-1L,
130  5.3357020033841180906486E-1L,
131  5.2213689121370692017331E-1L,
132  5.1094857432705833910408E-1L,
133  5.0000000000000000000000E-1L,
134 };
135 static const long double B[17] = {
136  0.0000000000000000000000E0L,
137  2.6176170809902549338711E-20L,
138 -1.0126791927256478897086E-20L,
139  1.3438228172316276937655E-21L,
140  1.2207982955417546912101E-20L,
141 -6.3084814358060867200133E-21L,
142  1.3164426894366316434230E-20L,
143 -1.8527916071632873716786E-20L,
144  1.8950325588932570796551E-20L,
145  1.5564775779538780478155E-20L,
146  6.0859793637556860974380E-21L,
147 -2.0208749253662532228949E-20L,
148  1.4966292219224761844552E-20L,
149  3.3540909728056476875639E-21L,
150 -8.6987564101742849540743E-22L,
151 -1.2327176863327626135542E-20L,
152  0.0000000000000000000000E0L,
153 };
154 
155 /* 2^x = 1 + x P(x),
156  * on the interval -1/32 <= x <= 0
157  */
158 static const long double R[] = {
159  1.5089970579127659901157E-5L,
160  1.5402715328927013076125E-4L,
161  1.3333556028915671091390E-3L,
162  9.6181291046036762031786E-3L,
163  5.5504108664798463044015E-2L,
164  2.4022650695910062854352E-1L,
165  6.9314718055994530931447E-1L,
166 };
167 
168 #define MEXP (NXT*16384.0L)
169 /* The following if denormal numbers are supported, else -MEXP: */
170 #define MNEXP (-NXT*(16384.0L+64.0L))
171 /* log2(e) - 1 */
172 #define LOG2EA 0.44269504088896340735992L
173 
174 #define F W
175 #define Fa Wa
176 #define Fb Wb
177 #define G W
178 #define Ga Wa
179 #define Gb u
180 #define H W
181 #define Ha Wb
182 #define Hb Wb
183 
184 static const long double MAXLOGL = 1.1356523406294143949492E4L;
185 static const long double MINLOGL = -1.13994985314888605586758E4L;
186 static const long double LOGE2L = 6.9314718055994530941723E-1L;
187 static const long double huge = 0x1p10000L;
188 /* XXX Prevent gcc from erroneously constant folding this. */
189 static const volatile long double twom10000 = 0x1p-10000L;
190 
191 static long double reducl(long double);
192 static long double powil(long double, int);
193 
powl(long double x,long double y)194 long double powl(long double x, long double y)
195 {
196 	/* double F, Fa, Fb, G, Ga, Gb, H, Ha, Hb */
197 	int i, nflg, iyflg, yoddint;
198 	long e;
199 	volatile long double z=0;
200 	long double w=0, W=0, Wa=0, Wb=0, ya=0, yb=0, u=0;
201 
202 	/* make sure no invalid exception is raised by nan comparision */
203 	if (isnan(x)) {
204 		if (!isnan(y) && y == 0.0)
205 			return 1.0;
206 		return x;
207 	}
208 	if (isnan(y)) {
209 		if (x == 1.0)
210 			return 1.0;
211 		return y;
212 	}
213 	if (x == 1.0)
214 		return 1.0; /* 1**y = 1, even if y is nan */
215 	if (x == -1.0 && !isfinite(y))
216 		return 1.0; /* -1**inf = 1 */
217 	if (y == 0.0)
218 		return 1.0; /* x**0 = 1, even if x is nan */
219 	if (y == 1.0)
220 		return x;
221 	if (y >= LDBL_MAX) {
222 		if (x > 1.0 || x < -1.0)
223 			return INFINITY;
224 		if (x != 0.0)
225 			return 0.0;
226 	}
227 	if (y <= -LDBL_MAX) {
228 		if (x > 1.0 || x < -1.0)
229 			return 0.0;
230 		if (x != 0.0 || y == -INFINITY)
231 			return INFINITY;
232 	}
233 	if (x >= LDBL_MAX) {
234 		if (y > 0.0)
235 			return INFINITY;
236 		return 0.0;
237 	}
238 
239 	w = floorl(y);
240 
241 	/* Set iyflg to 1 if y is an integer. */
242 	iyflg = 0;
243 	if (w == y)
244 		iyflg = 1;
245 
246 	/* Test for odd integer y. */
247 	yoddint = 0;
248 	if (iyflg) {
249 		ya = fabsl(y);
250 		ya = floorl(0.5 * ya);
251 		yb = 0.5 * fabsl(w);
252 		if( ya != yb )
253 			yoddint = 1;
254 	}
255 
256 	if (x <= -LDBL_MAX) {
257 		if (y > 0.0) {
258 			if (yoddint)
259 				return -INFINITY;
260 			return INFINITY;
261 		}
262 		if (y < 0.0) {
263 			if (yoddint)
264 				return -0.0;
265 			return 0.0;
266 		}
267 	}
268 	nflg = 0; /* (x<0)**(odd int) */
269 	if (x <= 0.0) {
270 		if (x == 0.0) {
271 			if (y < 0.0) {
272 				if (signbit(x) && yoddint)
273 					/* (-0.0)**(-odd int) = -inf, divbyzero */
274 					return -1.0/0.0;
275 				/* (+-0.0)**(negative) = inf, divbyzero */
276 				return 1.0/0.0;
277 			}
278 			if (signbit(x) && yoddint)
279 				return -0.0;
280 			return 0.0;
281 		}
282 		if (iyflg == 0)
283 			return (x - x) / (x - x); /* (x<0)**(non-int) is NaN */
284 		/* (x<0)**(integer) */
285 		if (yoddint)
286 			nflg = 1; /* negate result */
287 		x = -x;
288 	}
289 	/* (+integer)**(integer)  */
290 	if (iyflg && floorl(x) == x && fabsl(y) < 32768.0) {
291 		w = powil(x, (int)y);
292 		return nflg ? -w : w;
293 	}
294 
295 	/* separate significand from exponent */
296 	x = frexpl(x, &i);
297 	e = i;
298 
299 	/* find significand in antilog table A[] */
300 	i = 1;
301 	if (x <= A[17])
302 		i = 17;
303 	if (x <= A[i+8])
304 		i += 8;
305 	if (x <= A[i+4])
306 		i += 4;
307 	if (x <= A[i+2])
308 		i += 2;
309 	if (x >= A[1])
310 		i = -1;
311 	i += 1;
312 
313 	/* Find (x - A[i])/A[i]
314 	 * in order to compute log(x/A[i]):
315 	 *
316 	 * log(x) = log( a x/a ) = log(a) + log(x/a)
317 	 *
318 	 * log(x/a) = log(1+v),  v = x/a - 1 = (x-a)/a
319 	 */
320 	x -= A[i];
321 	x -= B[i/2];
322 	x /= A[i];
323 
324 	/* rational approximation for log(1+v):
325 	 *
326 	 * log(1+v)  =  v  -  v**2/2  +  v**3 P(v) / Q(v)
327 	 */
328 	z = x*x;
329 	w = x * (z * __polevll(x, P, 3) / __p1evll(x, Q, 3));
330 	w = w - 0.5*z;
331 
332 	/* Convert to base 2 logarithm:
333 	 * multiply by log2(e) = 1 + LOG2EA
334 	 */
335 	z = LOG2EA * w;
336 	z += w;
337 	z += LOG2EA * x;
338 	z += x;
339 
340 	/* Compute exponent term of the base 2 logarithm. */
341 	w = -i;
342 	w /= NXT;
343 	w += e;
344 	/* Now base 2 log of x is w + z. */
345 
346 	/* Multiply base 2 log by y, in extended precision. */
347 
348 	/* separate y into large part ya
349 	 * and small part yb less than 1/NXT
350 	 */
351 	ya = reducl(y);
352 	yb = y - ya;
353 
354 	/* (w+z)(ya+yb)
355 	 * = w*ya + w*yb + z*y
356 	 */
357 	F = z * y  +  w * yb;
358 	Fa = reducl(F);
359 	Fb = F - Fa;
360 
361 	G = Fa + w * ya;
362 	Ga = reducl(G);
363 	Gb = G - Ga;
364 
365 	H = Fb + Gb;
366 	Ha = reducl(H);
367 	w = (Ga + Ha) * NXT;
368 
369 	/* Test the power of 2 for overflow */
370 	if (w > MEXP)
371 		return huge * huge;  /* overflow */
372 	if (w < MNEXP)
373 		return twom10000 * twom10000;  /* underflow */
374 
375 	e = w;
376 	Hb = H - Ha;
377 
378 	if (Hb > 0.0) {
379 		e += 1;
380 		Hb -= 1.0/NXT;  /*0.0625L;*/
381 	}
382 
383 	/* Now the product y * log2(x)  =  Hb + e/NXT.
384 	 *
385 	 * Compute base 2 exponential of Hb,
386 	 * where -0.0625 <= Hb <= 0.
387 	 */
388 	z = Hb * __polevll(Hb, R, 6);  /*  z = 2**Hb - 1  */
389 
390 	/* Express e/NXT as an integer plus a negative number of (1/NXT)ths.
391 	 * Find lookup table entry for the fractional power of 2.
392 	 */
393 	if (e < 0)
394 		i = 0;
395 	else
396 		i = 1;
397 	i = e/NXT + i;
398 	e = NXT*i - e;
399 	w = A[e];
400 	z = w * z;  /*  2**-e * ( 1 + (2**Hb-1) )  */
401 	z = z + w;
402 	z = scalbnl(z, i);  /* multiply by integer power of 2 */
403 
404 	if (nflg)
405 		z = -z;
406 	return z;
407 }
408 
409 
410 /* Find a multiple of 1/NXT that is within 1/NXT of x. */
reducl(long double x)411 static long double reducl(long double x)
412 {
413 	long double t;
414 
415 	t = x * NXT;
416 	t = floorl(t);
417 	t = t / NXT;
418 	return t;
419 }
420 
421 /*
422  *      Positive real raised to integer power, long double precision
423  *
424  *
425  * SYNOPSIS:
426  *
427  * long double x, y, powil();
428  * int n;
429  *
430  * y = powil( x, n );
431  *
432  *
433  * DESCRIPTION:
434  *
435  * Returns argument x>0 raised to the nth power.
436  * The routine efficiently decomposes n as a sum of powers of
437  * two. The desired power is a product of two-to-the-kth
438  * powers of x.  Thus to compute the 32767 power of x requires
439  * 28 multiplications instead of 32767 multiplications.
440  *
441  *
442  * ACCURACY:
443  *
444  *                      Relative error:
445  * arithmetic   x domain   n domain  # trials      peak         rms
446  *    IEEE     .001,1000  -1022,1023  50000       4.3e-17     7.8e-18
447  *    IEEE        1,2     -1022,1023  20000       3.9e-17     7.6e-18
448  *    IEEE     .99,1.01     0,8700    10000       3.6e-16     7.2e-17
449  *
450  * Returns MAXNUM on overflow, zero on underflow.
451  */
452 
powil(long double x,int nn)453 static long double powil(long double x, int nn)
454 {
455 	long double ww, y;
456 	long double s;
457 	int n, e, sign, lx;
458 
459 	if (nn == 0)
460 		return 1.0;
461 
462 	if (nn < 0) {
463 		sign = -1;
464 		n = -nn;
465 	} else {
466 		sign = 1;
467 		n = nn;
468 	}
469 
470 	/* Overflow detection */
471 
472 	/* Calculate approximate logarithm of answer */
473 	s = x;
474 	s = frexpl( s, &lx);
475 	e = (lx - 1)*n;
476 	if ((e == 0) || (e > 64) || (e < -64)) {
477 		s = (s - 7.0710678118654752e-1L) / (s +  7.0710678118654752e-1L);
478 		s = (2.9142135623730950L * s - 0.5 + lx) * nn * LOGE2L;
479 	} else {
480 		s = LOGE2L * e;
481 	}
482 
483 	if (s > MAXLOGL)
484 		return huge * huge;  /* overflow */
485 
486 	if (s < MINLOGL)
487 		return twom10000 * twom10000;  /* underflow */
488 	/* Handle tiny denormal answer, but with less accuracy
489 	 * since roundoff error in 1.0/x will be amplified.
490 	 * The precise demarcation should be the gradual underflow threshold.
491 	 */
492 	if (s < -MAXLOGL+2.0) {
493 		x = 1.0/x;
494 		sign = -sign;
495 	}
496 
497 	/* First bit of the power */
498 	if (n & 1)
499 		y = x;
500 	else
501 		y = 1.0;
502 
503 	ww = x;
504 	n >>= 1;
505 	while (n) {
506 		ww = ww * ww;   /* arg to the 2-to-the-kth power */
507 		if (n & 1)     /* if that bit is set, then include in product */
508 			y *= ww;
509 		n >>= 1;
510 	}
511 
512 	if (sign < 0)
513 		y = 1.0/y;
514 	return y;
515 }
516 #elif LDBL_MANT_DIG == 113 && LDBL_MAX_EXP == 16384
517 // TODO: broken implementation to make things compile
powl(long double x,long double y)518 long double powl(long double x, long double y)
519 {
520 	return pow(x, y);
521 }
522 #endif
523